A MACHINE LEARNING-BASED READINESS FORECASTING MODEL FOR TRAINING LOAD DECISION SUPPORT

Abstract

This study proposes and evaluates a machine learning-based approach for forecasting athletes’ readiness one week ahead to support coaching decisions on training-load management. A secondary analysis was conducted using an open dataset from a 12-week controlled study involving 48 competitive table tennis players. Weekly monitoring data included training load, session RPE, ACWR, training monotony and strain, heart-rate variability, resting heart rate, sleep, fatigue, stress, soreness, wellness, and current readiness.

Four forecasting models - Elastic Net, Random Forest, XGBoost, and a multilayer perceptron - were evaluated using athlete-grouped nested cross-validation. Raw monitoring indicators were compared with personalised temporal features incorporating week-to-week changes, three-week moving averages, and athlete-specific deviations. Model interpretation was performed using out-of-fold SHAP analysis, and temporal validation was used to assess generalisation to later phases of the training programme.

The best monitoring-based model was Random Forest with personalised temporal features, achieving MAE = 0.835, RMSE = 1.039, and R2=0.152. Differences between algorithms were small, while personalised temporal features provided only modest improvement over raw indicators.

The results show that weekly monitoring data can provide a modest but interpretable forecast of next-week readiness. The proposed approach can support coaches by identifying athletes at risk of reduced readiness while preserving professional judgement in training-load decisions.